Check a julia-backend fit against a particle-filter likelihood
Source:R/ctParticleLik.R
ctParticleLik.RdThe filter ctsem fits with is an assumed-density approximation: it carries a Gaussian for the latent states, linearises the drift over each prediction step, and projects non-Gaussian measurement updates back onto a Gaussian. For a linear model with Gaussian indicators it is exact; for anything else the size of its error is invisible from inside it. This function estimates the same marginal log likelihood at the fit's estimates with a bootstrap particle filter that makes none of those approximations, and reports the difference, in total and per row.
Arguments
- fit
A fit from
ctFit(..., backend = 'julia')with the defaultintoverpop; random effects carried as augmented states are integrated by the particles. Laplace-route fits are refused.- particles
Particles per subject.
- substeps
Transition steps per observation interval. This is the reference mesh, independent of anything the fit used, and floored by the fit's own
maxtimestepor automatic mesh.- transition
'exponential'applies the filter's own discretisation at each particle's state: exact for the linear parts of the model, second order otherwise, so the comparison isolates the filter's Gaussian assumption.'euler'is plain Euler-Maruyama, which shares no approximation with the filter at all but needs several times more substeps.- seed
Fixes every draw; the result is reproducible given it.
- replicates
Independent runs (with seeds
seed,seed + 1, ...). One run reports an approximate standard error from the effective sample sizes; more than one reports the standard deviation across runs.- cores
Engine threads. The particle filter splits the subjects across them, each subject with its own random stream, so the result does not depend on the count.
Value
A list: loglik, the particle estimate (mean over replicates);
se; fit_loglik, the filter's log likelihood from the fit;
difference, particle minus filter; rows, a data frame with
the subject, time, particle and filter log likelihood of every row and
their difference; replicates, the individual estimates;
ess_min, the smallest effective sample size any row saw; and the
settings.
Details
Each particle steps through the process with drift, diffusion and every
state-dependent parameter evaluated at its own state, is weighted by the
conditional density of the row's observations given that state (Gaussian,
binary, ordinal, count and censored indicators are all handled), and is
resampled when the weights degenerate. The estimate is unbiased with Monte
Carlo error falling as one over the square root of particles.
It is a diagnostic at fixed parameters, not an estimator. Two things decide
whether to trust the comparison: se should be well below the
difference you care about (raise particles, or set replicates
for an empirical standard error), and doubling substeps should not
move the answer.
See also
ctFitUncertainty with method = 'is' and
filter = 'particle' applies the correction this
measures, to the fit's posterior draws.
Examples
# \donttest{
# Needs a working julia backend, so the example is inert where Julia is
# absent -- including on CRAN, whose check machines have none.
if (isTRUE(ctJuliaStatus()$available)) {
model <- ctModel(type = 'ct', manifestNames = 'Y1', latentNames = 'eta1',
LAMBDA = matrix(1))
model$pars$indvarying <- FALSE
fit <- ctFit(ctstantestdat, model, backend = 'julia', cores = 1)
# A linear Gaussian model, where the filter is exact: the difference
# should be within a few standard errors of zero.
check <- ctParticleLik(fit, particles = 1000, substeps = 5)
check[c('loglik', 'se', 'fit_loglik', 'difference')]
# Julia keeps running between calls, holding what it has compiled;
# this ends it.
ctJuliaKill("session")
}
# }